The Pudding, a digital publication devoted to data-driven visualization of current culture, currently features a very interesting essay on Wine and Math, A Model Pairing. The author, Lars Verspohl, provides many eye-catching graphics of the analytics behind producing quality wines.
What got my attention was a simulator for making red Portuguese Vinho Verde. Verspohl sifted through a dataset of 1600 wines to develop a model that predicts quality based on 11 factors. You can slide these up and down to try making a fine wine—rated at 7 or more on a scale of 10.
Not being content with haphazard searching on so many variables, I set up a multifactor test. Using version 13 of Design-Expert® software (free trial here), I laid out a minimum-run (plus 2) screening design on the 8 factors ranked most important by Verspohl’s Random Forest analysis, bypassing the bottom 3 (pH, residual sugar and free sulfur dioxide). I then worked through the 18 combinations and recorded the quality results in percent.
As shown on its Pareto plot of effects, Design-Expert revealed that only 5 of the effects tested produced significant effects.
The numeric optimization tools led to the optimal red Vinho Verde flagged in this 3D plot at the highest alcohol and lowest volatile acid levels. Settings for the other attributes are indicated by the position of the slide bars, e.g.; sulphates at the high level*. The factors defaulted to the middle are ones that did not get picked for the model.
Now that I’ve solved this simulator, my next mission is to locate a bottle of red Vinho Verde for some one-glass-at-a-time testing.
*This result surprised me—not being a big fan of sulfurous compounds in wines. This skepticism is borne out by another take on the Vinho Verde wine here. The only way to resolve the conflicting results would be to do an actual experiment on the composition of a red wine, ideally a mixture design for optimal formulation.